to be avoided and the function is pretty small.
3% speedup, though this is probably due to changed inlining and not directly
this change.

Originally committed as revision 16301 to svn://svn.ffmpeg.org/ffmpeg/trunk